The clearest shift in U.S. markets in September 2026 was not that AI stocks kept climbing. It was the widening gap between bonds and equities.
U.S. Treasury data showed the 10-year Treasury yield rose from 4.75% on Aug. 31 to 5.29% on Sept. 30, up 54 basis points in a single month. QuantStreet said some U.S. fixed-income assets fell 2.3% to 5% over the same period. Yet technology shares, especially semiconductors, stayed firm. Bitcoin, momentum strategies with major holdings in tech and chip stocks, and the Nasdaq all held up relatively well.
Elsewhere, the picture was weaker. U.S. small caps, the equal-weight S&P 500, and rate-sensitive groups including real estate investment trusts, utilities, and financials were broadly under pressure. The dollar and commodities also rose at the same time in September, an unusual combination given that they often move in opposite directions.
In his latest monthly investment letter, QuantStreet Capital founder Harry Mamaysky framed the divergence around two questions: why did technology stocks remain resilient when long-term rates moved sharply higher, and if investors are betting on AI-led growth, why has that view not spread more broadly across the stock market?
What was the market pricing as yields surged?
A move of more than 50 basis points in the 10-year yield over one month points to a meaningful repricing of long-term capital. Bond prices and yields usually move in opposite directions, and longer-duration securities tend to take the largest hit when rates rise.
Still, there is no single explanation that has won broad agreement for the Treasury sell-off.
One argument is that investors were losing confidence in the U.S. dollar. Mamaysky was not convinced. The dollar actually appreciated in September, which does not fit a story of broad-based selling in dollar assets. That does not rule out long-term credit concerns, but it does suggest the market was not dealing with a one-way crisis of confidence in the currency.
A second explanation is that the U.S. fiscal position was forcing investors to demand more long-term risk compensation. Mamaysky again stopped short of endorsing that conclusion. He pointed to relatively stable inflation breakevens, which did not rise nearly as much as nominal Treasury yields.
That measure captures the yield gap between nominal Treasuries and Treasury Inflation-Protected Securities, and it is often used to track how investors price future inflation and related risk compensation. If nominal yields are rising while breakevens are fairly stable, it becomes harder to attribute the entire move to runaway inflation expectations.
That said, stable breakevens do not eliminate fiscal risk. Long-term yields can also be shaped by Treasury supply, term premium, real rates, and market liquidity. By itself, this metric does not prove the absence of debt risk.
Mamaysky paid more attention to a third possibility: the market may be repricing for stronger future growth and for the heavy capital needs tied to AI infrastructure buildout.
Large cloud providers and technology companies have continued to expand capital expenditures on data centers, chips, and related power infrastructure. Those plans tie up more capital and can lift financing needs. If investors also expect AI to raise productivity and future profits, then higher long-term yields may reflect not only worsening risk conditions but also changing expectations for growth and capital demand.
He did not argue that AI investment has been proven to be the main driver of higher Treasury yields. His point was narrower: it may help explain why some technology stocks kept rising even as bonds sold off.
Why were AI stocks stronger in a higher-rate market?
Equity valuation can be thought of as the discounted value of future cash flows. All else equal, higher interest rates mean investors demand a higher rate of return, which lowers the present value of future earnings. That is why expensive growth stocks are usually more sensitive to rising rates.
September did not fully follow that script. Mamaysky’s explanation was that investors may believe AI will generate enough future earnings growth to offset the pressure from a higher discount rate.
In valuation terms, two forces were working against each other. A higher discount rate pulls down the present value of future profits, while stronger expected earnings lift intrinsic value. If the second force is large enough, stocks can continue to rise even with rates moving higher. In that reading, the market was not ignoring higher rates. It was assuming future AI profits could cover the higher cost of capital.
There is a practical basis for that view. AI infrastructure spending is already creating substantial demand. Chipmakers, semiconductor equipment suppliers, and related technology firms can capture revenue directly from the buildout. AMD, Micron, Intel, Cisco, and Applied Materials were among the major holdings in the momentum ETF Mamaysky discussed. As long as investors believe AI capital spending will remain elevated, earnings expectations for upstream suppliers may stay supported.
But that does not settle the larger question. Revenue growth at those companies starts with higher capital spending by other firms. It does not automatically mean the wider economy has already realized matching productivity gains.
For companies buying chips and building data centers, the spending first appears as cost or as a capital asset. The investment only becomes a durable economic return if those assets eventually lift revenue, reduce costs, or improve profitability.
That is why a rally in semiconductor shares can show confidence in AI infrastructure names without proving that the ultimate economic payoff from AI has already arrived.
The central gap: chip companies are earning, but what about everyone else?
This was the tension Mamaysky emphasized most.
Semiconductors were strong in September, while the equal-weight S&P 500 was weak. Unlike a market-cap-weighted index, the equal-weight version gives each constituent roughly the same weight, making it a better gauge of whether gains are spreading across the market rather than being driven by a small group of large technology stocks.
Mamaysky referred to the broader set of companies outside semiconductors as ROCS, short for the Rest of the Corporate Sector.
In his view, AI investment follows a chain that takes time to verify. Companies are buying chips, servers, and software today because they expect these tools to raise productivity later. Capital markets are willing to finance that buildout in advance because they expect profits that have not yet been realized.
On that basis, it is not surprising that profits are not rising in lockstep across all industries right now. But stock markets are forward-looking. If investors were convinced that AI would significantly improve future earnings for the broader corporate sector, some of that expectation should gradually appear in those share prices as well.
That broad repricing did not happen in September. Semiconductor stocks kept advancing, while other companies did not receive similar valuation support. Mamaysky’s question followed naturally: if the end buyers of chips cannot earn enough incremental profit, how can they keep absorbing the rising cost of AI purchases and infrastructure buildout over time?
The issue goes straight to profit allocation across the AI investment chain. In the near term, infrastructure suppliers may earn strong profits through order growth and tight demand. Over the medium term, cloud providers need to recover their spending by leasing compute capacity and selling AI services. Over a longer horizon, ordinary companies need to convert AI into higher productivity, lower operating costs, or new revenue streams.
Only if that process starts to materialize can the value created by AI investment spread through broader economic activity.
At the same time, weak price action outside semiconductors does not prove that productivity gains from AI will never show up. High interest rates, company-specific operating pressure, and different valuation starting points can all mask expectations for future earnings improvement. Even so, September’s trading suggested that investor confidence in upstream AI suppliers was clearly stronger than confidence in the eventual beneficiaries across the rest of the corporate sector.
Mamaysky did not conclude that AI investment is bound to fail. He said he still believes in AI’s long-term value and does not see the current market as a bubble. His point was that this rally still needs broader fundamental support, and that means profits generated by AI must extend beyond a narrow group of technology firms.
What comes next: productivity, profits, and portfolio positioning
As he evaluates AI’s economic value, Mamaysky has started paying closer attention to productivity data.
Revised data released by the U.S. Bureau of Labor Statistics on Sept. 3, 2026 showed nonfarm business labor productivity rose at an annualized quarter-over-quarter rate of 1.4% in the second quarter and 2.2% year over year. Over a longer span, from the fourth quarter of 2019 through the second quarter of 2026, nonfarm business labor productivity increased at an annual average rate of about 2.1%, above roughly 1.5% in the prior business cycle. Mamaysky said that fit the improvement trend he had been observing.
He also drew a line between a macro productivity pickup and confirmed AI contribution. Labor productivity can be affected by capital investment, labor allocation, technological progress, and cyclical factors. It cannot yet be used to calculate directly how much profit AI has created for companies.
That led to a stricter test: productivity improvement ultimately has to show up in corporate earnings outside the technology sector.
The same view has influenced QuantStreet’s portfolio changes. Even though value stocks and low-volatility stocks underperformed in the previous quarter, the firm kept a relative overweight in both areas to retain exposure to the broader corporate sector. In portfolios with higher risk tolerance, it also kept some exposure to technology stocks.
In fixed income, QuantStreet has started adjusting duration. Duration measures how sensitive a bond’s price is to changes in yield. Longer duration usually means larger price losses when yields rise, but also more upside if yields fall.
Mamaysky said the investment appeal of bonds began to improve once the 10-year Treasury yield reached about 5.25%. Based on that view, QuantStreet slightly increased bond duration in lower-risk portfolios, marking one of the firm’s more noticeable directional changes in more than a year.
That did not amount to a broad bullish call on long-duration bonds. The firm’s models still do not favor high-duration fixed-income assets, and overall portfolio duration remains below benchmark, though the underweight is now smaller.
For suitable investors, Mamaysky also mentioned the diversification role of alternative assets such as evergreen private equity funds. Based on product performance figures he cited, some of those funds rose about 0.5% to 0.75% in September, offering some portfolio diversification in a month when most equities faced pressure. He also noted the limits: valuation frequency, liquidity, and underlying asset risk remain important constraints, and one month of returns does not prove long-term defensiveness.
Those adjustments show that QuantStreet has not exited the AI trade, nor has it made a large rotation into long bonds simply because yields have moved higher. The firm is trying to balance risk and return across assets instead.
The real test ahead comes down to three signals.
- First, whether AI infrastructure spending can continue and whether revenue growth at upstream suppliers still has strong enough demand behind it.
- Second, whether AI begins to improve profitability for non-tech companies. Productivity data can offer an early clue, but margins, cost savings, and new revenue are more direct evidence that investment returns are being realized.
- Third, whether the rise in long-term Treasury yields reflects growth expectations more than inflation, fiscal supply, and term-risk compensation. If growth fails to improve as expected while financing costs stay high, pressure on investment returns will increase.
The AI trade now faces a broader question than how long demand for chips and compute can keep rising. It has to answer how much incremental profit these investments can create for the wider economy.
Semiconductor companies have already booked visible revenue from capital spending. Broader earnings improvement across the corporate sector remains unproven. Only when productivity gains from AI begin to turn into real profit outside the technology sector will markets have fuller evidence to support the scale of current investment.